Machine Learning for Exoplanet Detection in High-Contrast Spectroscopy: Revealing Exoplanets by Leveraging Hidden Molecular Signatures in Cross-Correlated Spectra with Convolutional Neural Networks

TL;DR

The aim of this method is to leverage weak assumptions on exoplanet characterisation, such as the presence of specific molecules in atmospheres, to improve detection sensitivity for exoplanets.

Abstract

The new generation of observatories and instruments (VLT/ERIS, JWST, ELT) motivate the development of robust methods to detect and characterise faint and close-in exoplanets. Molecular mapping and cross-correlation for spectroscopy use molecular templates to isolate a planet's spectrum from its host star. However, reliance on signal-to-noise ratio (S/N) metrics can lead to missed discoveries, due to strong assumptions of Gaussian independent and identically distributed noise. We introduce machine learning for cross-correlation spectroscopy (MLCCS); the method aims to leverage weak assumptions on exoplanet characterisation, such as the presence of specific molecules in atmospheres, to improve detection sensitivity for exoplanets. MLCCS methods, including a perceptron and unidimensional convolutional neural networks, operate in the cross-correlated spectral dimension, in which patterns from molecules can be identified. We test on mock datasets of synthetic planets inserted into real noise from SINFONI at K-band. The results from MLCCS show outstanding improvements. The outcome on a grid of faint synthetic gas giants shows that for a false discovery rate up to 5%, a perceptron can detect about 26 times the amount of planets compared to an S/N metric. This factor increases up to 77 times with convolutional neural networks, with a statistical sensitivity shift from 0.7% to 55.5%. In addition, MLCCS methods show a drastic improvement in detection confidence and conspicuity on imaging spectroscopy. Once trained, MLCCS methods offer sensitive and rapid detection of exoplanets and their molecular species in the spectral dimension. They handle systematic noise and challenging seeing conditions, can adapt to many spectroscopic instruments and modes, and are versatile regarding atmospheric characteristics, which can enable identification of various planets in archival and future data.

Figures (25)

• Figure 1: Molecular maps of H$_2$O for real PZ Tel B data using cross-correlation for spectroscopy. This figure shows a real case example where the noise structures may reduce detection capabilities of cross-correlation methods. The brown dwarf was observed under good conditions (airmass: $1.11$, Seeing start to end: $0.77-0.72$) and lower conditions (airmass: $1.12$, Seeing: $1.73-1.54$), c.f. Appendix \ref{['supp:molmap']} for full details on observing conditions. Upper plots show molecular maps of PZ Tel B, while the lower plots show the cross-correlation series along the radial velocity (RV) support for pixels at the centre of the object, and within the object's brightness area. While the brown dwarf should appear at the same spatial coordinates for respective RV locations in both cases (c.f. vertical lines), it is clearly visible when conditions are good, but hardly visible on equal scales under lower conditions.
• Figure 2: Flowchart representing the methods, scoring and classification workflows presented across sections. Each cross-correlated spatial pixel (spaxel) is treated as an independent instance and is passed through a classifier of static (statistic) or a dynamic (learning algorithm) type. The methods will evaluate the RV series and yield scoring metrics (e.g. a statistic or probability score). In order to perform classification, the scores need to first be separated using a meaningful threshold. The current standard classification scheme is yield by the signal-to-noise ratio (S/N) on the cross-correlation peak at the planet's RV. We propose to analyse the RV series in a holistic approach using machine learning (ML) to detect the planets and their molecules, and use the resulting probability scores.
• Figure 3: Illustration of the shape and size of one cross-correlated dataframe. Each row is a sample, and it represents a cross-correlated spaxel (CC spaxel). The RV steps are called features, and the elements of the last column Y is the categorical labelling indicating the presence of a planet or molecule of interest in the spectra. One whole cross-correlated dataframe as above is named a template channel; it results from the cross-correlation of the whole spectral dataset (i.e. all samples) with a unique template.
• Figure 4: Architecture of the perceptron. This simple one-layer neural network analyses the values of a whole cross-correlation series in a holistic approach, to detect the presence of a molecular signal.
• Figure 5: Example of the application of a Convolutional Neural Network (CNN) on one cross-correlated spectrum. For each sample of a cross-correlation N, and for all M template channels, the convolution filter runs across the channels and along the RV series. The filter depth is M, and its size is optimised according to the training. Hence, for a same series cross-correlated with M different templates, those convolutional layers allow to filter out important and recurrent patterns.